What v0.2 did
sequence (or uploaded PDB) -> 3D structure prediction RhoFold+ (single-sequence default) -> conformational ensemble 5 frames, ANM normal-mode sampling (ProDy) -> cavity detection per frame fpocket, RNA-tuned parameters -> cross-frame clustering at 4 A Kabsch alignment, persistence-aware -> ranking persistence x binding-residue stability -> top-3 shortlist + PDB + JSON + PDF
RhoFold+ was used under its Apache-2.0 software licence, weights for inference only and never fine-tuned, its training data being non-commercial. ProDy (BSD-3) provided the normal-mode analysis and fpocket (MIT) the cavity detection. Those attributions stand; it is the interpretation built on top of them that did not.
What was later measured against it
Claim
A top-3 ranked shortlist of candidate druggable pockets.
Method
The ranking used fpocket’s druggability score.
The control that disproved it
Measured directly against real RNA binding sites, that score consistently returns at or near zero for the true site while rewarding cavities that are not binding sites. It is a learned function trained on protein pockets. The recovery rates reported by v0.2 are survivorship over a ranker that is inverted on this molecule class.
What runs now
Pocket detection is reported as detection and geometry. The druggability score is discarded on nucleic acids and used on protein, where it is valid.
Claim
A five-frame ANM conformational ensemble broadens pocket detection.
Method
Anisotropic network model normal-mode sampling around one predicted structure.
The control that disproved it
ANM perturbs a structure along its own normal modes, so a cleft present in the starting model is preserved across every frame. Cavities that are artefacts of the prediction survive the ensemble rather than being filtered by it, and persistence across frames is therefore a partial mitigation and not a test.
What runs now
Multiple independent engines replace the single-model ensemble. Two engines disagreeing about a cavity is evidence of a different kind from one engine agreeing with a perturbed copy of itself.
Claim
Cross-target scoring separates promising targets from unpromising ones.
Method
Scoring metrics applied across targets to prioritise them.
The control that disproved it
No metric tested — pose convergence, interface confidence, contact agreement — separated real binders from decoys on RNA. A docking result of 0.778 AUC fell to 0.506 once an apo control was run. A published pocket-scoring model scoring 0.647 on its own pocket scored 0.633 on unrelated pockets. Eight approaches were tested and eight failed the same way.
What runs now
No composite score, no target-quality score and no go/no-go verdict is produced anywhere in the platform.
Claim
Model confidence identifies which predicted structure to trust.
Method
pLDDT used to select among replicates.
The control that disproved it
Across an apo replicate-selection study, no blind signal captured more than about 5% of the available accuracy, and pLDDT performed worse than choosing at random. Separately, a 90-compound benchmark ran on eighteen structures that contained none of the target’s defining feature while every replicate agreed with every other.
What runs now
Confidence is reported exactly as the engine emits it and is never used to rank or select. Model-independent structural checks run first instead.
What survived
Not all of it was wrong, and the parts that held are in the current pipeline.
- fpocket with the RNA-tuned parameters from Veenbaas et al. (PNAS 2025) genuinely does detect cavities better on nucleic acids than protein defaults do — positive predictive value 78% against 19%. Detection was never the problem; ranking was.
- Reporting the residues within 5 Å of a cavity centre, rather than a single score, is still how pockets are described.
- Full provenance on every result, and a downloadable bundle containing the parameters that produced it.
If you read the earlier version of this page and used it to form a view of a target, the specific numbers that matter are in the controls above. We would rather tell you than let the page quietly change underneath you. Questions to info@rnafold.com.